Fujitsu Limited
Developing explainable AI technology that solves problems in the field of genomic medicine, such as classifying cancer types, with the world’s highest accuracy
Deriving cause and effect by automatically combining judgment materials in different data formats such as text and images
……
Our company recently announced that it uses AI to automatically integrate and learn from multiple different formats of data such as text, images, and numbers into a knowledge graph that abstracts and systematizes knowledge. We have developed explainable AI technology that determines and estimates the cause and content with high accuracy. To confirm the effectiveness of this technology, we tested it on multiple benchmarks, including challenges in the medical field, such as classifying lung cancer types and predicting survival times for breast cancer patients. As a result, we confirmed that, for example, it is possible to support the classification of the two main types of lung cancer with high accuracy by tracing back to
pathological image information and explaining the factors.
We have also developed a technology that extracts and learns features from images with completely different ways of depicting objects, such as paintings, line drawings, illustrations, and photographs, and makes highly accurate judgments. By combining these technologies, it is expected that it will be possible to create AI that supports highly accurate decisions by combining genomic information, even when sufficient training data such as pathological images cannot be prepared.
In the future, we will proceed with the development of these developed multimodal technologies (Note 1) as general-purpose technologies that can be applied not only in the medical field but also in various fields, and by the end of fiscal 2024, we will be able to try out various cutting-edge technologies developed by our company. It is scheduled to be published on the Fujitsu Research Portal (Note 2). 【background】
In recent years, AI has spread rapidly and is used in a variety of situations, but most of it is specialized for specific data formats or learning models. However, in the real world, there are many situations in which decisions are made by simultaneously considering multiple perspectives on a single issue. For example, in the medical field, more accurate judgments can be made by integrating and analyzing multiple data such as test numerical information, image information, and genomic information, and AI can also be used in such situations. Although it is expected to be used, it has been difficult to integrate and learn from multiple fields, fields, and data formats.
[About the development technology]
Our company is conducting research and development on multimodal technology that handles data from different fields and formats, and has recently developed a new technology that integrates and learns from multiple different data formats and image data with completely different ways of depicting objects. We have developed the following two AI technologies that make integrated judgments from various perspectives.
1. AI technology that integrates and learns image data in which objects are depicted in completely different ways, such as line drawings and photographs.
By integrating and learning data from images in which objects are depicted in different ways, such as paintings, line drawings, illustrations, and photographs, it is possible to make appropriate judgments regarding objects based on new input image data such as photographs. We have developed a technology that can do this. In new practical situations, situations often arise in which it is difficult to obtain a sufficient amount of input image data, and this technology will contribute to solving this problem. This method learns both features specific to the way the object is drawn and features common to the object regardless of how the object is drawn. This makes it possible to learn knowledge and make appropriate judgments regarding objects even if the objects are drawn in various ways, such as pictures, line drawings, and illustrations, in the learning data. This technology is standardly used in this research field and uses three benchmarks (PACS, Office-Home, and DomainNet) that have datasets of multiple types of images with different ways of depicting objects such as art, manga, and photographs. When the target object was judged, an accuracy improvement of around 2% was observed compared to conventional technology. These results were accepted at ICLR 2024, a prestigious academic conference called The International Conference on Learning Representations (ICLR) (Note 3), and our company made a presentation on this matter on May 8, 2024.
2. Explainable AI technology that integrates data in different formats and transforms it into a common knowledge graph for learning. The above technology is extended not only to different image data but also to different data formats such as text and images, converting each data into a common knowledge graph independent of format, and then automatically integrating them using AI. We created a large-scale integrated knowledge graph and developed a new technology that uses it to support decisions in an explainable manner.
When this technology was applied to the following medical fields, we were able to achieve performance that exceeds that of conventional technology. In the future, this technology will be combined with AI technology that integrates and learns image data that depicts objects in completely different ways, and will be able to support
decision-making by combining genomic information even when making judgments on pathological images for which sufficient learning data cannot be prepared. It is hoped that you will be able to do so. 1. Type classification of lung cancer Treatment methods are being established for each type of lung cancer, such as adenocarcinoma (Note 4) and squamous cell carcinoma (Note 5), and correct treatment requires accurate classification. It is important. Until now, doctors had to visually check multiple pieces of information during
examinations, but this technology now uses AI to automatically collect pathological images and genome information (copy number abnormality information) of lung cancer patients. We integrated them and separated them into cancer types. As a result, an evaluation using data from The Cancer Genome Atlas (TCGA), which is commonly used as a benchmark around the world, revealed that the highest accuracy in classifying lung cancer types was 87.1%, whereas this technology achieved 92.1%. Achieved the world’s highest accuracy. When classifying these types, it is possible to refer back to pathological image data to provide the basis for the judgment.
2. Determining survival prediction for breast cancer patients If patients can accurately predict survival time for each treatment when choosing a treatment method, they will be more likely to select an appropriate treatment method. This time, in addition to image data of breast cancer patients, RNA data (Note 6) and clinical data are automatically integrated and judged using AI, and evaluated using benchmark data from The Cancer Genome Atlas (TCGA). In the task of predicting the survival period of breast cancer patients, the highest accuracy was previously 66.8%, but this technology achieved an accuracy of 71.8%. When supporting these predictions of survival time, it is possible to demonstrate the basis from image data.
[Image 1: https://prtimes.jp/i/93942/294/resize/d93942-294-ee243726ea8db88ef8d4-0.png&s3=93942-294-8330384027c4c203e308f08871fac5fc-980×376.png ]
Figure 1 Conversion into a common graph format that integrates data in different formats
【About the future】
The technology we have developed is scheduled to be made public in 2024 on the Fujitsu Research Portal, an environment where you can try out various cutting-edge technologies developed by our company. In 2023, the company will begin joint research with the Spanish research institute Barcelona Supercomputing Center (Centro Nacional de Supercompputacion, hereinafter referred to as the Barcelona
Supercomputing Center) in the area of personalized medicine. We plan to utilize the AI technology developed using this multimodal technology in joint research with the Barcelona Supercomputing Center, aiming to further improve accuracy and gain global recognition. Furthermore, we will continue to develop this technology with a view to using it not only in the medical field, but also in a variety of other fields, such as data center failure prediction and fraud detection.
[About trademark]
Proper nouns such as product names listed are trademarks or registered trademarks of each company.
[Note]
Note 1
Multimodal technology: A technology that handles multiple data in different formats such as text, images, and numerical values.In the medical field, it is a technology that can handle patient information in electronic medical records, test results, CT images, genome databases, etc. in a unified manner. About.
Note 2
Fujitsu Research Portal: This portal site will be open to the public from June 2023, allowing trial use of Fujitsu’s cutting-edge technology by registering an account.
Note 3
The International Conference on Learning Representations (ICLR): The world’s top international conference on representation learning (deep learning)
Note 4
Adenocarcinoma: A type of lung cancer that develops from epithelial tissue called glandular tissue, which is part of the body’s tissues. It is the most common type of lung cancer, accounting for about half of all cases.
Note 5
Squamous cell carcinoma: A type of lung cancer that develops from the mucous tissue called squamous epithelium, which is one of the tissues that make up the body. It is the second most common type of lung cancer after adenocarcinoma, accounting for approximately 20-30% of all lung cancers.
Note 6
RNA data: RNA is a substance that is the source of proteins that make up living things, and RNA data is data that can be processed by computers using sequencers.
Note 7
Barcelona Supercomputing Center: Location Barcelona, Spain Director Mateo Valero
【Related Links】
Fujitsu Research Portal (https://portal.research.global.fujitsu.com/) [About our contribution to SDGs]
[Image 2: https://prtimes.jp/i/93942/294/resize/d93942-294-671a5cc54dcdfc7cb7c2-1.png&s3=93942-294-c9a168709863c6dbe278e18fd5bca756-196×112.png ]
The Sustainable Development Goals (SDGs) adopted by the United Nations in 2015 are common goals that the entire world should achieve by 2030. Our purpose (raison d’être), “To bring trust to society and make the world more sustainable through innovation,” is our promise to contribute to the SDGs.
[Image 3: https://prtimes.jp/i/93942/294/resize/d93942-294-7e33ab56ea402776c68d-2.png&s3=93942-294-90594a35e38a948e5700d2e5ac1a01fe-577×185.png ]
Main SDGs that this project aims to contribute to
[Inquiries regarding this matter]
Fujitsu Contact Line (General Counter)
0120-933-200 (toll free)
Reception hours: 9:00 to 12:00 and 13:00 to 17:30 (excluding Saturdays, Sundays, holidays, and Fujitsu designated holidays) Contact form (https://contactline.jp.fujitsu.com/customform/csque04802/873532/) Product prices, specifications, service details, etc. listed in the press release are as of the date of announcement. It is subject to change without prior notice. Please note.
More details about this release:
https://prtimes.jp/main/html/rd/p/000000294.000093942.html